Abstract Air pollution is reckoned as one of the main critical factors in risks related to human health as well as to the environment, especially in key areas like the Arctic, where global warming is affecting the ecosystem at an alarming rate. Major pollutants, such as Particulate Matter (PM 10 ), are thus crucial to monitor for efficient planning ahead of extreme events: to that extent, a reliable short to medium-term forecast is vital for prompt action. We propose a Deep Learning approach to air pollution forecasting aimed at predicting 48-h concentrations of PM 10 . We analyse the behaviour of said pollutant together with auxiliary factors, such as meteorological variables and outputs from state-of-the-art numerical models, through several forefront time-series-specialised architectures, i.e. Transformers. We develop an enhanced version of the best-performing Transformer architecture to define a suitable model to forecast PM 10 concentrations from a wide variety of stations across Northern Europe. Our proposed methodology manages to outperform pivotal numerical models and associated postprocessing, offering a valuable alternative for local air pollution forecasting.
Cuzzucoli et al. (Wed,) studied this question.